Method and system for detecting pesticide residues in fruits and vegetables based on spectral attention network

Abstract

The application discloses a fruit and vegetable pesticide residue detection method and system based on a spectral attention network, first, fruit and vegetable hyperspectral images are collected, and data preprocessing is performed; then, the preprocessed hyperspectral images are input into the spectral attention network for pesticide residue detection; the application can automatically detect pesticide residues in food, has the advantages of high detection efficiency, high accuracy, simple operation and the like, and provides an effective means for food safety supervision.

Description

Technical Field

[0001] This invention belongs to the field of food safety testing technology, and relates to a method and system for detecting pesticide residues in fruits and vegetables, and more particularly to a method and system for detecting pesticide residues in fruits and vegetables based on spatial spectral attention networks. Background Technology

[0002] Methods for pesticide residue detection mainly include spectroscopic methods (Reference 1), enzyme inhibition methods (Reference 2), and chromatographic methods (Reference 3). However, these methods still have many problems, such as high cost, long processing time, and low efficiency (Reference 4). Therefore, hyperspectral technology has become a highly anticipated technique, capable of obtaining high-resolution continuous spectral data and detecting pesticide residues by reflecting the internal physicochemical properties of the target. Using hyperspectral technology for pesticide residue detection is convenient and efficient, and does not damage the target (References 5-9). However, traditional hyperspectral pesticide residue detection methods require manual feature design and cannot extract the complex nonlinear spatial-spectral features of hyperspectral images, resulting in insufficient detection accuracy.

[0003] D.Liu, Y.Han, L.Zhu, W.Chen, Y.Zhou, J.Chen, and Z.Dou, "Quantitative detection of isofenphos-methyl in corns using surface-enhanced Ramanspectroscopy (SERS) with chemometric methods," Food Analytical Methods, vol.10, no.5, pp.1202-1208, 2017.

[0004] HOU X, SHEN G. Study on elimination of false positives in the detection of spicy vegetables by wheat gluten esterase inhibition method [J]. Food and Fermentation Industries, 47(4):247-252.

[0005] Xu Bingwen, Sun Chengpeng, Ge Xiangwu, et al. Determination of matrix effect and its overcoming mechanism of 61 pesticides in 30 kinds of fruits and vegetables by gas chromatography-tandem mass spectrometry [J]. Journal of Food Safety and Quality Inspection, 2021, 12(15):6068-6076.

[0006] Li Bo, Song Shanshan, Xia Jianfeng, et al. Application of modern food safety testing technology in the detection of residues in agricultural products [J]. Food Research and Development, 2020, (14): 127-132.

[0007] Wang Miaozhi, Jiang Hongjun, Sun Hongyang. Application of hyperspectral imaging technology in pesticide residue detection in fruits and vegetables [J]. Research on Agricultural Modernization, 2019, (4): 23-25.

[0008] Wang,Q.,Wang,X.,Zeng,W.,&Wang,C.(2019).Hyperspectral imaging technology in food quality and safety detection: a review.Critical reviews infood science and nutrition,59(2),253-271.

[0009] Li,P.,Cao,K.,Zhang,L.,&He,Y.(2020).A review of hyperspectral imaging technology in food safety and quality detection.Journal of Food Protection, 83(4),718-735.

[0010] Chen,Y.,Huang,Y.,&Wang,J.(2020).Pesticide residue detection based on hyperspectral technology:research progress and perspective.Journal of integrated agriculture, 19(5),1000-1014.

[0011] Bao,Y.,Li,Y.,Zhang,C.,&Zhou,J.(2020).Non-destructive detection of pesticide residues in agricultural products based on hyperspectral imaging technology: a review.Food Analytical Methods, 13(5),1012-1030. Summary of the Invention

[0012] To address the problems of low detection accuracy, high cost, long detection time, and low efficiency in existing detection methods, this invention provides a method and system for detecting pesticide residues in fruits and vegetables based on spatial spectrum attention networks.

[0013] The technical solution adopted by the method of the present invention is: a method for detecting pesticide residues in fruits and vegetables based on spatial spectral attention networks, comprising the following steps:

[0014] Step 1: Acquire hyperspectral images of fruits and vegetables and perform data preprocessing;

[0015] Step 2: Input the preprocessed hyperspectral image into the spatial-spectral attention network for pesticide residue detection;

[0016] The spatial-spectral attention network includes an encoder module, a decoder module, a spatial attention module, a spectral attention module, and a classifier module;

[0017] The encoder module includes four convolutional layers arranged in series. Each of the first, second, and third convolutional layers is followed by a residual block, and each convolutional layer is followed by a normalization layer and an activation layer.

[0018] The decoder module includes four convolutional layers arranged in series. After the first, second, and third convolutional layers, a feature fusion block is added to sum the output features from the first, second, and third convolutional layers of the encoder and decoder, respectively. After each convolutional layer, a normalization layer and an activation layer are added in sequence.

[0019] The spatial attention module and the spectral attention module are arranged in parallel between the encoder module and the decoder module;

[0020] The classifier module, located after the decoder module, includes a fully connected layer and a softmax activation function, used to transform the feature map into a one-dimensional vector and then map it to the class label distribution.

[0021] Preferably, in step 1, the preprocessing includes denoising and spectral normalization; firstly, the original spectral data is smoothed by filtering to remove high-frequency noise points, improve the signal-to-noise ratio, and smooth the original data sequence. Then, the hyperspectral data is normalized to eliminate differences between different spectral data.

[0022] Preferably, in step 2, the encoder module includes four convolutional layers arranged in series, each of which is a convolutional layer with a kernel size of 3 and a stride of 1.

[0023] The decoder module includes four convolutional layers arranged in series, each of which is a convolutional layer with a kernel size of 3 and a stride of 1;

[0024] The spatial attention module, for each pixel, sets its feature vector f ij With three learnable weight matrices, W respectively v Multiply to obtain the query vector qij Key vector k ij Sum vector v ij Then, the attention weight matrix A is obtained by performing a dot product operation and a softmax operation between the query vector and the key vector of each pixel; finally, the spatial attention module multiplies the attention weight matrix A with the value matrix V of all pixels to obtain the final output feature map.

[0025] For an input feature map X, the spectral attention module first performs a global average pooling operation, compressing the feature map of each channel into a scalar. For each channel, the spectral attention module introduces two learnable weight matrices W. s and W e The spectral attention module compresses the scalar z of each channel. i respectively with W s and W e Multiplying them together yields a scaling factor s. i and an offset e i The spectral attention module will analyze the feature map X of each channel. i Multiply by the scaling factor s respectively i Then add the offset e i The weighted feature map Y is obtained. i Ultimately, the spectral attention module applies all weighted feature maps Y... i By concatenating these features, we obtain the output feature map Y.

[0026] Preferably, in step 2, the spatial spectrum attention network is a trained spatial spectrum attention network; the training process includes the following sub-steps:

[0027] Step S1: Prepare fruit and vegetable samples and collect their hyperspectral data;

[0028] Step S2: Preprocess the acquired hyperspectral image data;

[0029] Step S3: Create a pesticide concentration detection dataset;

[0030] Step S4: Input the preprocessed hyperspectral data into the spatial-spectral attention network for training, extract the spectral and spatial features related to pesticide residues, and continuously optimize the model parameters through the backpropagation algorithm so that it can accurately identify pesticide residues.

[0031] As a preferred option, in step S1, a number of fruits and vegetables are collected, washed and dried to ensure that no pesticide residue remains on the skin, and then pesticide diluted with water is sprayed evenly on the surface of the fruits and vegetables; hyperspectral data are collected using an imaging spectrometer.

[0032] Preferably, in step S2, the preprocessing includes denoising and spectral normalization; firstly, the original spectral data is smoothed by filtering to remove high-frequency noise points, improve the signal-to-noise ratio, and smooth the original data sequence. Then, the hyperspectral data is normalized to eliminate differences between different spectral data.

[0033] Preferably, in step S3, the process of creating a pesticide concentration detection dataset involves first labeling the hyperspectral data to obtain pesticide concentration labels for different sample data, and then dividing the labeled dataset into a training set and a test set.

[0034] Preferably, in step S4, the cross-entropy loss function is used during training until the network converges, meaning the training loss curve remains stable and no longer decreases. The pesticide concentration level with the highest predicted probability is taken as the final detection result.

[0035] The technical solution adopted by the system of the present invention is: a fruit and vegetable pesticide residue detection system based on spatial spectrum attention network, comprising:

[0036] One or more processors;

[0037] A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the fruit and vegetable pesticide residue detection method based on spatial spectral attention network.

[0038] This invention can automatically detect pesticide residues in fruits and vegetables, and has the advantages of high detection efficiency, high accuracy and simple operation, providing an effective means for the safety supervision of fruits and vegetables. Attached Figure Description

[0039] The technical solutions described herein are further illustrated below using examples and specific implementation methods. Additionally, accompanying drawings are used in the description of the technical solutions. Those skilled in the art can, without any creative effort, obtain other drawings and the intent of the present invention based on these drawings.

[0040] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0041] Figure 2 This is a diagram of the spatial spectrum attention network structure according to an embodiment of the present invention;

[0042] Figure 3 This is a flowchart of the training process of the spatial spectrum attention network according to an embodiment of the present invention. Detailed Implementation

[0043] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0044] Please see Figure 1 This invention provides a method for detecting pesticide residues in fruits and vegetables based on a spatial spectral attention network, comprising the following steps:

[0045] Step 1: Acquire hyperspectral images of fruits and vegetables and perform data preprocessing;

[0046] In one implementation, preprocessing includes denoising and spectral normalization; firstly, the original spectral data is smoothed by filtering to remove high-frequency noise points, improve the signal-to-noise ratio, and smooth the original data sequence. Then, the hyperspectral data is normalized to eliminate differences between different spectral data.

[0047] Step 2: Input the preprocessed hyperspectral image into the spatial-spectral attention network for pesticide residue detection;

[0048] Please see Figure 2 The spatial-spectral attention network in this embodiment includes an encoder module, a decoder module, a spatial attention module, a spectral attention module, and a classifier module;

[0049] The encoder module of this embodiment includes four convolutional layers arranged in series, each of which is a convolutional layer with a kernel size of 3 and a stride of 1; residual blocks are added after the first, second and third convolutional layers, and a normalization layer and an activation layer are added sequentially after each convolutional layer.

[0050] The decoder module in this embodiment includes four convolutional layers arranged in series, each of which is a convolutional layer with a kernel size of 3 and a stride of 1; after the first, second and third convolutional layers, a feature fusion block is added to sum the output features from the first, second and third convolutional layers of the encoder and decoder, respectively; after each convolutional layer, a normalization layer and an activation layer are added in sequence.

[0051] In this embodiment, the spatial attention module, for each pixel, assigns its feature vector f... ij With three learnable weight matrices, W respectively v Multiply to obtain the query vector q ij Key vector k ij Sum vector v ij Then, the attention weight matrix A is obtained by performing a dot product operation and a softmax operation between the query vector and the key vector for each pixel. Finally, the spatial attention module multiplies the attention weight matrix A with the value matrix V of all pixels to obtain the final output feature map.

[0052] In this embodiment, the spectral attention module, for an input feature map X, first performs a global average pooling operation, compressing the feature map of each channel into a scalar. For each channel, the spectral attention module introduces two learnable weight matrices W. s and W e These are used to calculate a scaling factor and an offset, respectively. Specifically, the spectral attention module compresses the scalar z for each channel. i respectively with W s and W e Multiplying them together yields a scaling factor s. i and an offset e i Scaling factor s i The feature map used to weight each channel, and the offset e i Used to adjust the weighted results. Specifically, the spectral attention module adjusts the feature map X of each channel. i Multiply by the scaling factor s respectively i Then add the offset e i The weighted feature map Y is obtained. i Ultimately, the spectral attention module processes all weighted feature maps Y i The concatenations yield the output feature map Y;

[0053] The classifier module in this embodiment includes a fully connected layer and a softmax activation function, which are used to transform the feature map into a one-dimensional vector and then map it to the class label distribution.

[0054] Please see Figure 3 The spatial spectrum attention network implemented in this paper is a pre-trained spatial spectrum attention network; the training process includes the following sub-steps:

[0055] Step S1: Prepare fruit and vegetable samples and collect their hyperspectral data.

[0056] In one implementation, Jiaozhou Chinese cabbage is used as an example of fruit and vegetable samples, but the method is applicable to other vegetables, fruits, meats, and aquatic products. Four commonly used commercially available pesticides for Chinese cabbage were diluted with water and sprayed onto the surface of the cabbage, including abamectin, imidacloprid, acetamiprid, and carbendazim. Hyperspectral data was acquired using the MateSpec LS portable hyperspectral imager, which is a staring imaging system developed based on a line-scanning pixel-coated hyperspectral sensor and has a built-in scanning hyperspectral imaging system. The wavelength range is 470-900 nm, and the device weighs only 0.8 kg.

[0057] The specific spectral acquisition process of the portable high-resolution imaging spectrometer is as follows:

[0058] (1) Turn on the spectrometer and preheat for 15 minutes;

[0059] (2) Establish a connection between the imaging spectrometer and the computer;

[0060] (3) Set the spectral measurement parameters;

[0061] (4) Calibrate the instrument using a reference standard or standard sample;

[0062] (4) Click the start button to collect spectral data;

[0063] (5) Export the collected data for further analysis and processing.

[0064] Step S2: Preprocess the acquired hyperspectral image data.

[0065] In one implementation, the hyperspectral data is first imported into Matlab. Different noise removal methods are selected based on the specific data and the problem. If the noise is significant, a filter is used for denoising. If the noise is complex, wavelet transform is chosen. Wavelet transform decomposes the original hyperspectral data sequence into different frequency bands, and noise is removed based on the characteristics of each band. After noise removal, the smoothed original hyperspectral data sequence is subjected to min-max normalization, mapping the data to the 0-1 range. This involves subtracting the minimum value from each data point and dividing by the difference between the maximum and minimum values. Normalization unifies data from different bands into a single range, facilitating subsequent data processing and analysis.

[0066] Step S3: Create a hyperspectral pesticide concentration detection dataset.

[0067] In one implementation, pesticide concentrations from different sample hyperspectral data are labeled and combined into a dataset. Each data point contains the sample's hyperspectral data and the corresponding residual pesticide concentration label. The entire dataset is divided into a training set and a test set at a ratio of 80% / 20%.

[0068] Step S4: Input the preprocessed hyperspectral data into a spatial-spectral attention network for training. Extract spectral and spatial features related to pesticide residues. Continuously optimize the model parameters through backpropagation to accurately identify pesticide residues. After model training is complete, evaluate the model's accuracy and generalization performance by validating it on a test set.

[0069] In one implementation, the spatial-spectral attention network structure consists of an input layer, convolutional layer 1, pooling layer 1, convolutional layer 2, pooling layer 2, a fully connected layer, and an output layer. The convolutional layers incorporate BatchNormalization and ReLU activation functions. The former normalizes the data in each layer of the neural network, thereby accelerating training and avoiding problems such as vanishing or exploding gradients. The latter increases the nonlinearity of the neural network, accelerating training while improving the model's prediction accuracy. Preprocessed hyperspectral data is input into the spatial-spectral attention network model for training, continuously extracting spectral and spatial features through multiple convolutional and pooling layers. During training, the backpropagation algorithm is typically used to continuously adjust the model parameters, minimizing the model's loss function. After training, the model's accuracy and generalization performance can be evaluated by validating it on a test set. Once the model has sufficient accuracy and robustness, it can be used to detect unknown samples and determine whether the pesticide concentration meets the required standards.

[0070] The spatial-spectral attention network in this embodiment includes an encoder module, a decoder module, a spatial attention module, a spectral attention module, and a classifier module;

[0071] The encoder module of this embodiment includes four convolutional layers arranged in series, each of which is a convolutional layer with a kernel size of 3 and a stride of 1; residual blocks are added after the first, second and third convolutional layers, and a normalization layer and an activation layer are added sequentially after each convolutional layer.

[0072] The decoder module in this embodiment includes four convolutional layers arranged in series, each of which is a convolutional layer with a kernel size of 3 and a stride of 1; after the first, second and third convolutional layers, a feature fusion block is added to sum the output features from the first, second and third convolutional layers of the encoder and decoder, respectively; after each convolutional layer, a normalization layer and an activation layer are added in sequence.

[0073] In this embodiment, the spatial attention module, for each pixel, assigns its feature vector f... ij With three learnable weight matrices, W respectively v Multiply to obtain the query vector q ij Key vector k ij Sum vector v ij Then, the attention weight matrix A is obtained by performing a dot product operation and a softmax operation between the query vector and the key vector for each pixel. Finally, the spatial attention module multiplies the attention weight matrix A with the value matrix V of all pixels to obtain the final output feature map.

[0074] In this embodiment, the spectral attention module, for an input feature map X, first performs a global average pooling operation, compressing the feature map of each channel into a scalar. For each channel, the spectral attention module introduces two learnable weight matrices W. s and W e These are used to calculate a scaling factor and an offset, respectively. Specifically, the spectral attention module compresses the scalar z for each channel. i respectively with W s and W e Multiplying them together yields a scaling factor s. i and an offset e i Scaling factor s i The feature map used to weight each channel, and the offset e i Used to adjust the weighted results. Specifically, the spectral attention module adjusts the feature map X of each channel. i Multiply by the scaling factor s respectively i Then add the offset e i The weighted feature map Y is obtained. i Ultimately, the spectral attention module processes all weighted feature maps Y i The concatenations yield the output feature map Y;

[0075] The classifier module in this embodiment includes a fully connected layer and a softmax activation function, which are used to transform the feature map into a one-dimensional vector and then map it to the class label distribution.

[0076] The introduction of a non-linear activation function in this embodiment enables deep neural networks to learn non-linear features, significantly enhancing the network's expressive power. Commonly used activation functions include the Sigmoid function, Softmax function, Tanh function, ReLU function, and Leaky-ReLU function. Among these, the ReLU non-linear activation function has a simple form, effectively avoids the gradient vanishing problem, and accelerates model convergence. Its specific form is as follows:

[0077]

[0078] The cross-entropy loss function is used during training until the network converges, meaning the training loss curve remains stable and no longer decreases. The pesticide concentration level with the highest predicted probability is taken as the final detection result.

[0079] In one implementation, when inputting new sample data into a trained model for pesticide residue detection, the data needs to be preprocessed to adapt it to the model. Common preprocessing operations include normalization, scaling, and cropping. Next, the preprocessed data is fed into the model for forward computation to obtain the detection results. Further analysis of the detection results determines whether pesticide residues exceed the limits, thereby ensuring food safety.

[0080] This invention can automatically detect pesticide residues in food, and has the advantages of high detection efficiency, high accuracy and simple operation, providing an effective means for food safety supervision.

[0081] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for detecting pesticide residues in fruits and vegetables based on a spatial spectral attention network, characterized in that, Includes the following steps: Step 1: Acquire hyperspectral images of fruits and vegetables and perform data preprocessing; The preprocessing includes denoising and spectral normalization; firstly, the original spectral data is smoothed by filtering to remove high-frequency noise points, improve the signal-to-noise ratio, and smooth the original data sequence; then, the hyperspectral data is normalized to eliminate the differences between different spectral data. Step 2: Input the preprocessed hyperspectral image into the spatial-spectral attention network for pesticide residue detection; The spatial-spectral attention network includes an encoder module, a decoder module, a spatial attention module, a spectral attention module, and a classifier module; The encoder module includes four convolutional layers arranged in series. Each of the first, second, and third convolutional layers is followed by a residual block, and each convolutional layer is followed by a normalization layer and an activation layer. The decoder module includes four convolutional layers arranged in series. After the first, second, and third convolutional layers, a feature fusion block is added to sum the output features from the first, second, and third convolutional layers of the encoder and decoder, respectively. After each convolutional layer, a normalization layer and an activation layer are added in sequence. The spatial attention module and the spectral attention module are arranged in parallel between the encoder module and the decoder module; The classifier module, located after the decoder module, includes a fully connected layer and a softmax activation function, used to transform the feature map into a one-dimensional vector and then map it to the class label distribution; The spatial spectrum attention network is a trained spatial spectrum attention network; The training process includes the following sub-steps: Step S1: Prepare fruit and vegetable samples and collect their hyperspectral data; Step S2: Preprocess the acquired hyperspectral image data; Step S3: Create a pesticide concentration detection dataset; Step S4: Input the preprocessed hyperspectral data into the spatial-spectral attention network for training, extract the spectral and spatial features related to pesticide residues, and continuously optimize the model parameters through the backpropagation algorithm to enable it to accurately identify pesticide residues; the cross-entropy loss function is used during training until the network converges, that is, the training loss curve remains stable and no longer decreases; the pesticide concentration level with the highest predicted probability is taken as the final detection result.

2. The method for detecting pesticide residues in fruits and vegetables based on spatial spectrum attention network according to claim 1, characterized in that: In step 2, the encoder module includes four convolutional layers arranged in series, each of which is a convolutional layer with a kernel size of 3 and a stride of 1; The decoder module includes four convolutional layers arranged in series, each of which is a convolutional layer with a kernel size of 3 and a stride of 1; The spatial attention module, for each pixel, sets its feature vector f ij With three learnable weight matrices, W respectively v Multiply to obtain the query vector q ij Key vector k ij Sum vector v ij Then, the attention weight matrix A is obtained by performing a dot product operation and a softmax operation between the query vector and the key vector of each pixel; finally, the spatial attention module multiplies the attention weight matrix A with the value matrix V of all pixels to obtain the final output feature map. ; For an input feature map X, the spectral attention module first performs a global average pooling operation, compressing the feature map of each channel into a scalar. For each channel, the spectral attention module introduces two learnable weight matrices W. s and W e The spectral attention module compresses the scalar z of each channel. i respectively with W s and W e Multiplying them together yields a scaling factor s. i and an offset e i The spectral attention module will analyze the feature map X of each channel. i Multiply by the scaling factor s respectively i Then add the offset e i The weighted feature map Y is obtained. i Ultimately, the spectral attention module applies all weighted feature maps Y... i By concatenating these features, we obtain the output feature map Y.

3. The method for detecting pesticide residues in fruits and vegetables based on spatial spectrum attention network according to claim 1, characterized in that: In step S1, a number of fruits and vegetables are collected and washed to ensure that no pesticide residue remains on the skin. Then, pesticide diluted with water is sprayed evenly on the surface of the fruits and vegetables. Hyperspectral data is collected using an imaging spectrometer.

4. The method for detecting pesticide residues in fruits and vegetables based on spatial spectrum attention network according to claim 1, characterized in that: In step S3, the pesticide concentration detection dataset is created by first labeling the hyperspectral data to obtain pesticide concentration labels in different sample data, and then dividing the labeled dataset into training set and test set.

5. A pesticide residue detection system for fruits and vegetables based on a spatial spectrum attention network, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method for detecting pesticide residues in fruits and vegetables based on a spatial spectral attention network as described in any one of claims 1 to 4.

Images